Robust deepfake detector against deep image watermarking.

Area of Science:

• Computer Science
• Information Security
• Artificial Intelligence

Background:

• Deepfake technology presents a growing threat to information security.
• Existing deepfake detection methods often fail when images contain deep watermarks.
• Watermarking techniques like MBRS and FaceSigns can degrade detection performance.

Purpose of the Study:

• To develop a robust deepfake detection model resistant to image watermarking.
• To improve the accuracy of deepfake detection in the presence of common watermarking algorithms.

Main Methods:

• Proposed a multi-module deepfake detection model.
• Integrated Efficient Multi-scale Attention within the Xception architecture.
• Introduced a feature dropout module to remove redundant image features.

Main Results:

• The model achieved comparable accuracy to baseline models with MBRS watermarks.
• The model significantly outperformed baseline models with FaceSigns watermarks, showing 10% and 20% higher accuracy at 50% and 100% watermark presence, respectively.
• The feature dropout module effectively eliminated redundant image features.

Conclusions:

• The proposed model demonstrates enhanced robustness against deep watermarking in deepfake images.
• The integration of Efficient Multi-scale Attention and feature dropout improves detection performance, particularly against FaceSigns watermarks.
• This research contributes to more reliable deepfake detection systems in real-world scenarios.